dicom-series-preflight

This NVIDIA MedTech skill performs a header-only preflight check on a single DICOM series folder before that data is converted or fed to inference. It solves an early quality-gate problem in medical-imaging pipelines: catching malformed inventories, unexpected orientation, or potential PHI flags at the metadata level before spending time on pixel decoding or a full conversion. It deliberately does not decode pixel data, de-identify, or make any clinical determination.

The skill runs an upstream Python entrypoint, scripts/preflight_series.py, over a DICOM directory and emits JSON containing an inventory, orientation axcodes, PHI flags, findings, and an overall preflight verdict of pass, warn, or fail. It can be paired with a verifier and run through documented make workflows to produce a trusted preflight pack, and ships with fixtures, an output schema validator, tests, and a benchmark report. Users are instructed to read the skill_manifest.yaml before changing arguments and to use the wrapper exactly as documented rather than reimplementing it. Its published benchmark reports a passing NVSkills-Eval verdict with a 100% security dimension.

It targets medical-imaging and MLOps engineers building DICOM ingestion or model-inference pipelines who need a fast, read-only sanity gate. Stated limitations are important to the intended use: it is header-only and cannot detect burnt-in PHI, its canonical-orientation gate assumes LPS-derived CT axcodes, compressed and multi-frame instances are warned rather than decoded, and it scans a single directory. The documentation is explicit that it is not for clinical deployment, regulatory de-identification, autonomous diagnosis, or production ingestion without a vetted converter.